Is DeepSeek Helping to Solve the Energy Crisis?

During the recent Spring Festival holiday, families across China will have been discussing the latest global news stories around their dinner tables. Recent headlines have been dominated by the emergence of DeepSeek, a domestically developed AI platform that is being held up as a symbol of national achievement. DeepSeek’s notable edge comes from its ability to train large language models using much less computing power than its global competitors.

In the US, DeepSeek has already raised concerns among policymakers and sent shockwaves through the chipmaking industry. Share values of US energy producers and providers, including Constellation Energy, Vistra, and Talen Energy Corp, fell by 20 to 30 per cent on 27 January. This downturn reflects investor scepticism over the energy sector's decision to significantly expand power generation capacity, which is being justified as a necessary step to support AI innovation.

Nevertheless, industry analysts have noted that while DeepSeek’s training may be more energy-efficient, its reasoning model and extended responses could lead to higher energy consumption overall, compared to other generative AI models. Furthermore, there is another paradoxical trend at play; as AI model training becomes more efficient, it could encourage companies to invest even more money and energy in perfecting these models, leading to increased demand from end users. In simpler terms, using energy more efficiently can actually lead to a rise in total energy usage, a phenomenon referred to as the ‘Jevons paradox’.

In any case, the burgeoning AI industry will continue to create immense demand for data centres and energy sources. In China, a nation with ambitious objectives for both AI and sustainability, data centres consume more than 3 per cent of the country’s electricity. Last year, the central government released an action plan to accelerate the low-carbon transformation of China’s data centres, aiming to achieve an average power usage effectiveness (PUE) ratio below 1.5 by the end of 2025 – this means more than two-thirds of the total energy used should be spent directly on computing tasks. At the same time, a strategic plan is currently underway to build new data centres in western regions, where computing would be powered by wind, solar, and hydraulic power resources.